Evidence and claims

Summary

Biological analysis moves through different levels of evidence and inference. A measurement, statistical association, model prediction, mechanistic hypothesis, and causal conclusion are different objects and should not be collapsed into one another.

Core rules

A prediction is not a measurement.
Statistical association does not by itself establish molecular mechanism or causality.
Enrichment of signal in a set does not prove the predicted effect of every member of that set.
A model score must be interpreted according to its defined target, scale, and calibration.
Experimental validation should measure the quantity relevant to the claim being validated whenever practical.

Required context

For an evidence-based claim, identify where relevant:

what was directly measured
what was predicted or inferred
the unit and scale
the comparison being made
the model or statistical target
the experimental or observational design
uncertainty or confidence
whether the claim is associative, mechanistic, or causal

AI behaviour

Name the evidence type before escalating the interpretation.
Do not describe an association as a mechanism unless mechanistic evidence is available.
Do not describe a ranking score as a probability unless it is defined and calibrated as one.
Match validation evidence to the predicted quantity.

Common failure modes

Association presented as mechanism

A variant set that improves a burden-test association may contain more biologically relevant variants, but the association alone does not establish how an individual variant changes transcription, splicing, binding, or another molecular process.

Rank presented as probability

A percentile or PHRED-like rank can describe relative position among scored variants without representing the probability that a variant is pathogenic or functionally active.

Authoritative standards

Use the statistical, experimental, or clinical framework appropriate to the claim. Where a field has formal evidence criteria, cite that framework rather than translating it into an improvised local scale.

Examples

Model output

Prefer: “The model ranks this variant in the top 1% under its defined impact score.”

Avoid: “The variant has a 99% probability of being pathogenic” unless that probability is the validated quantity produced by the model.

Sources

Hernán MA, Robins JM. Causal Inference: What If. https://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/
National Academies. Reproducibility and Replicability in Science. https://doi.org/10.17226/25303